- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper introduces KOT, which uses source-modality dynamics and mechanistic relationships in the target modality to learn unpaired cross-modal mappings. The RNA–protein alignment and CRISPR protein-effect prediction experiments show promise. I find the direction meaningful, but it remains unclear which information drives the improvements and how the additional requirements affect the method’s practical use.
Critical strengths
1.The idea of incorporating dynamics into unpaired learning is clear and supported by controlled synthetic experiments.
2.The method achieves good alignment results on two CITE-seq datasets. A model trained only on non-targeting cells also retains predictive signal for protein effects under unseen perturbations.
3.The ablations are extensive. The authors’ discussion of the differences between static alignment, kinetic consistency, and perturbation prediction helps clarify the method’s capabilities and limitations.
Critical weaknesses
1.The contribution of dynamics to static alignment needs clarification. Shuffling, reversing, or even zeroing the velocity leaves alignment performance close to that of the full model, whereas disrupting the gene–protein correspondence substantially worsens it. This suggests that the static alignment gains may come largely from the structural prior. Comparing against a model with the entire kinetic loss removed does not sufficiently separate these contributions.
2.The perturbation results are encouraging, but the conclusions should remain limited. The model receives experimentally observed post-perturbation RNA, and the evaluation covers population-average effects for a small protein panel in one dataset. The advantage over a simple cognate-mRNA reference is also modest. These results support some transfer capability, but do not yet establish broadly accurate recovery of intervention responses.
3.The practical benefit of the unpaired setting is not yet fully demonstrated. The method removes the need for cell-level paired supervision, but still requires data suitable for velocity estimation, gene–protein relationships, and, in some experiments, half-life references. The paper could explain more concretely when these inputs are easier to obtain than paired measurements. The current benchmarks hide correspondence in originally paired datasets and therefore do not directly test the full workflow on independently collected populations. Runtime measurements on mixed hardware provide only a rough reference, and the cost of velocity preprocessing also needs to be considered.
Questions for authors
1.With the same gene–protein prior retained, how much additional benefit comes from using the correct dynamics? We would appreciate a clearer comparison that isolates the contribution of the dynamics model from that of the prior.
2.Can the perturbation results be reproduced across more proteins or datasets? Comparisons with methods given access to the same information would also help readers assess how broadly the reported improvements generalize.
3.Do preprocessing and model selection also avoid cell-level pairing information? Please clarify how RNA size-factor normalization is performed in Papalexi and whether FOSCTTM is used during model development, since these details affect the fairness of the comparison.
4.What practical setting do the authors consider best suited to KOT? Could you also report the total runtime, including velocity preprocessing, and peak memory usage on consistent hardware? This would help readers better assess the computational cost of using the method.